Merge branch 'main' into mlperf-inference

Author: arjunsuresh

Diff for: script/get-rocm-devices/README.md

@@ -0,0 +1,4 @@
+Run this script
+```
+cm run script --tags=get,rocm-devices 
+```

Diff for: script/get-rocm-devices/_cm.yaml

@@ -0,0 +1,29 @@
+alias: get-rocm-devices
+uid: c618239543364753
+
+automation_alias: script
+automation_uid: 5b4e0237da074764
+
+tags:
+- get
+- rocm-devices
+
+cache: false
+
+can_force_cache: true
+
+category: ROCM automation
+
+clean_files:
+- tmp-run.out
+docker:
+  run: false
+  all_gpus: 'yes'
+  skip_run_cmd: 'no'
+  skip_cm_sys_upgrade: 'yes'
+  cm_repo_flags: '--checkout=dev'
+  use_host_group_id: 'yes'
+  image_tag_extra: '-cm-dev'
+
+print_files_if_script_error:
+- tmp-run.out

Diff for: script/get-rocm-devices/customize.py

@@ -0,0 +1,62 @@
+from cmind import utils
+import os
+import subprocess
+
+def preprocess(i):
+
+    env = i['env']
+
+    if str(env.get('CM_DETECT_USING_HIP-PYTHON', '')).lower() in [ "1", "yes", "true"]:
+        i['run_script_input']['script_name'] = 'detect'
+
+    return {'return':0} 
+
+def postprocess(i):
+
+    env = i['env']
+    state = i['state']
+
+    os_info = i['os_info']
+
+    r = utils.load_txt(file_name='tmp-run.out',
+                       check_if_exists = True,
+                       split = True)
+    if r['return']>0: return r
+
+    lst = r['list']
+
+    # properties
+    p = {}
+    gpu = {}
+
+    gpu_id = -1
+
+    for line in lst:
+        #print (line)
+
+        j = line.find(':')
+
+        if j>=0:
+            key = line[:j].strip()
+            val = line[j+1:].strip()
+
+            if key == "GPU Device ID":
+                gpu_id+=1
+                gpu[gpu_id] = {}
+
+            if gpu_id < 0:
+                continue
+
+            gpu[gpu_id][key] = val 
+            p[key] = val
+
+            key_env = 'CM_ROCM_DEVICE_PROP_'+key.upper().replace(' ','_')
+            env[key_env] = val
+    
+    state['cm_rocm_num_devices'] = gpu_id + 1
+    env['CM_ROCM_NUM_DEVICES'] = gpu_id + 1
+
+    state['cm_rocm_device_prop'] = p
+    state['cm_rocm_devices_prop'] = gpu
+    
+    return {'return':0} 

Diff for: script/get-rocm-devices/detect.py

@@ -0,0 +1,55 @@
+from hip import hip
+
+# Defining the value for hipDeviceGetAttribute
+STRINGLENGTH = 256
+hipDeviceAttributeClockRate = 5
+hipDeviceAttributeMaxBlockDimX = 26
+hipDeviceAttributeMaxBlockDimY = 27
+hipDeviceAttributeMaxBlockDimZ = 28
+hipDeviceAttributeMaxGridDimX = 29
+hipDeviceAttributeMaxGridDimY = 30
+hipDeviceAttributeMaxGridDimZ = 31
+hipDeviceAttributeMaxThreadsPerBlock = 56
+hipDeviceAttributeMaxThreadsPerMultiProcessor = 57
+hipDeviceAttributeMaxRegistersPerBlock = 71
+hipDeviceAttributeMaxSharedMemoryPerBlock = 74
+hipDeviceAttributeWarpSize = 87
+
+
+def get_gpu_info():
+    num_gpus = hip.hipGetDeviceCount()[1]
+    all_gpu_info = []
+
+    for i in range(num_gpus):
+        gpu_info = {
+            "GPU Device ID": hip.hipDeviceGetPCIBusId(STRINGLENGTH, i)[1],
+            "GPU Name": i,
+            "GPU compute capability": f"{hip.hipDeviceComputeCapability(i)[1]}.{hip.hipDeviceComputeCapability(i)[2]}",
+            "ROCM driver version": f"{hip.hipDriverGetVersion()[1]}",
+            "ROCM runtime version": hip.hipRuntimeGetVersion()[1],
+            "Global memory (GiB)": hip.hipDeviceTotalMem(i)[1] / 1_073_741_824,
+            "Max clock rate": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeClockRate), i)[1] / 1000} MHz",
+            "Total amount of shared memory per block (Bytes)": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxSharedMemoryPerBlock), i)[1]}",
+            "Total number of registers available per block (Bytes)": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxRegistersPerBlock), i)[1]}",
+            "Warp size": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeWarpSize), i)[1]}",
+            "Maximum number of threads per multiprocessor": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxThreadsPerMultiProcessor), i)[1]}",
+            "Maximum number of threads per block": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxThreadsPerBlock), i)[1]}",
+            "Max dimension size of a thread block X": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxBlockDimX), i)[1]}",
+            "Max dimension size of a thread block Y": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxBlockDimY), i)[1]}",
+            "Max dimension size of a thread block Z": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxBlockDimZ), i)[1]}",
+            "Max dimension size of a grid size X": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxGridDimX), i)[1]}",
+            "Max dimension size of a grid size Y": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxGridDimY), i)[1]}",
+            "Max dimension size of a grid size Z": f"{hip.hipDeviceGetAttribute(hip.hipDeviceAttribute_t(hipDeviceAttributeMaxGridDimZ), i)[1]}",
+        }
+        all_gpu_info.append(gpu_info)
+
+    return all_gpu_info
+
+
+if __name__ == "__main__":
+    gpu_info_list = get_gpu_info()
+    with open("tmp-run.out", "w") as f:
+        for idx, gpu_info in enumerate(gpu_info_list):
+            print(f"GPU {idx}:")
+            for key, value in gpu_info.items():
+                f.write(f"{key}: {value}\n")

Diff for: script/get-rocm-devices/detect.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+${CM_PYTHON_BIN_WITH_PATH} ${CM_TMP_CURRENT_SCRIPT_PATH}/detect.py
+test $? -eq 0 || exit $?

Diff for: script/get-rocm-devices/run.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+# Compile
+
+rm a.out
+
+# Check if hip-python is installed
+echo ""
+echo "Checking if hip-python is installed..."
+echo ""
+
+if ! python3 -m pip show hip-python > /dev/null 2>&1; then
+    echo "hip-python not found. Installing hip-python..."
+    python3 -m pip install --extra-index-url https://test.pypi.org/simple hip-python
+    if [ $? -ne 0 ]; then
+        echo "Failed to install hip-python. Please check your Python environment."
+        exit 1
+    fi
+else
+    echo "hip-python is already installed."
+fi
+
+echo ""
+echo "Running program ..."
+echo ""
+
+cd ${CM_TMP_CURRENT_PATH}
+
+python ${CM_TMP_CURRENT_SCRIPT_PATH}/detect.py > tmp-run.out
+cat tmp-run.out
+test $? -eq 0 || exit 1

Diff for: script/get-tensorrt/customize.py

@@ -100,7 +100,7 @@ def preprocess(i):
     my_tar.close()
 
     import re
-    version_match = re.match(r'TensorRT-(\d+.\d+.\d+.\d+)', folder_name)
+    version_match = re.match(r'TensorRT-(\d+\.\d+\.\d+\.\d+)', folder_name)
     if not version_match:
         return {'return': 1, 'error': 'Extracted TensorRT folder does not seem proper - Version information missing'}
     version = version_match.group(1)